Experiments in Clustering Homogeneous XML Documents to Validate an Existing Typology

Transcription

1 Experiments in Clustering Homogeneous XML Documents to Validate an Existing Typology Thierry Despeyroux Yves Lechevallier Brigitte Trousse Anne-Marie Vercoustre (AxIS research team, Inria, France Abstract: This paper presents some experiments in clustering homogeneous XML documents to validate an existing classification or more generally an organisational structure. Our approach integrates techniques for extracting knowledge from documents with unsupervised classification (clustering) of documents. We focus on the feature selection used for representing documents and its impact on the emerging classification. We mix the selection of structured features with fine textual selection based on syntactic characteristics. We illustrate and evaluate this approach with a collection of Inria activity reports for the year The objective is to cluster projects into larger groups (Themes), based on the keywords or different chapters of these activity reports. We then compare the results of clustering using different feature selections, with the official theme structure used by Inria. Key Words: XML clustering, categorisation, organisational structure, knowledge discovery Category: H.3.1, I.5.3, I Introduction With the increasing amount of available information, sophisticated tools for supporting users in finding useful information are needed. In addition to tools for retrieving relevant documents, there is a need for tools that synthesise and exhibit information that is not explicitly contained in the document collection, using document mining techniques. Document mining objectives include extracting structured information from rough text, as well as document classification and clustering. Classification aims at associating documents to one or several predefined categories, while the objective of clustering is to identify emerging classes that are not known in advance. Traditional approaches for document classification and clustering rely on various statistical models, and representation of documents mostly based on bags of words. An important characteristic of text categorisation is the size of the vocabulary, which is often referred as the high dimension of the feature space. Automatic feature selection methods have been proposed to reduce the dimension of the space. They usually try to identify representative words that

2 can discriminate documents between various classes. For a comparison of those methods for classification see [Yang and Pederson 1997]. XML documents are becoming ubiquitous because of their rich and flexible format that can be used for a variety of applications. Standard methods have been used to classify XML documents, reducing them to their textual parts. These approaches do not take advantage of the structure of XML documents that also carries important information. In this paper we focus on XML documents and we study the impact of selecting (different) parts of the documents for a specific clustering task. The idea is that different parts of XML documents correspond to different dimensions of the collection that may play different roles in the classification task. We therefore consider two levels of feature selection: 1) coarse selection at the structure level and 2) fine linguistic selection of words within the text of elements. Based on the selected features the documents are then clustered using a dynamical classification algorithm that builds a prototype of each cluster as the union of all the features (words) of the documents belonging to this cluster. Furthermore, for each resulting cluster, we can exhibit the words that discriminate this cluster. Our experimentations use the collection of activity reports that were produced by the research groups at Inria. The task is to identify meaningful themes that would group projects working in related research domains. These groups are then compared to two Expert grouping, the first one used by Inria until 2003, and the new one proposed in Inria Activity Report Every year, Inria (The French National Institute for Research in Computer Science and Control) publishes an activity report (RA) made available to the French parliament and to our industrial and research partners. Traditionally produced as a paper document, this report is now published as a CD-Rom and the scientific part is made available on the Web since 1996 (in HTML and PDF). It is a collection of reports written by every Inria research team (in English since 2002). The XML version of these documents contains 139 files, a total of lines, more than 14.8 Mbytes of data. If the logical structure is defined by a DTD, the overall style and content are very flexible and unconstrained. The top level part of the DTD is given below: <!ELEMENT raweb (header, moreinfo?, members, presentation, foundation?, domain?, software?, results, contracts?, international?, dissemination?, biblio)> <!ATTLIST raweb year CDATA #IMPLIED > Mandatory sections include the list of team members, presentation of objectives, new results, and the list of publications for the year (biblio). Optional

3 sections include research foundation, application domains, software, as well as international and national cooperations. Inria research teams are also grouped into scientific themes that act as virtual structures for the purpose of presentation, communication and evaluation. The number of themes and allocation of teams to themes were decided some years ago by the board of directors and have changed recently. Choice of themes and team allocation are mostly related to strategic objectives and scientific closeness between existing teams. This has motivated our experiments in comparing the automatic clustering of teams, based on self-descriptions in their activity reports, with the two sets of themes defined by the Organisation. We will call them Expert Themes 2003 and 2004 respectively. Without anticipating on the results of the experiments, we are interested in discovering possible natural grouping of teams, identifying the keywords that better characterise those groups, and the potential difference with the organisational structures. 3 Methodology for Clustering XML Documents As said above, our objective is to cluster the research teams in themes, using their activity reports as data source. We hypothesis that activity reports reflect the research domains the teams are involved in and that some parts or the reports are more representative than others in describing research. For example, conferences and journals where researchers publish are indicative of their research fields. 3.1 Structure Selection and Vocabulary Definition The first step consists in selecting various parts of the XML documents that may be relevant for the classification task. This extraction uses the tools described in [Despeyroux 2004] to extract the text of elements, but standard XML tools could be used instead when the extraction does not require any inference. As we expect that various parts of the activity report would play different roles in classifying teams, we ran five experiments using different parts of the activity report, that are well-identified XML elements. We call this process structured feature selection. In this step, the documents are represented by the text of the selected elements. 1. Experiment K-F: Keywords attached to the foundation part 2. Experiment K-all: Keywords, whatever the sections they are attached to. 3. Experiment T-P: Full text of the presentation part 4. Experiment T-PF: Full text of the presentation and foundation parts 5. Experiment T-C: Names of conferences, workshops, congress, etc,

4 Experiences number extrated selected vocaof projects words words bulary K-F K-all T-P T-PF T-C Table 1: Size of data for the various experiments The goal of these experiments is to evaluate which parts are more relevant for the clustering task. The second processing step consists in the automatic selection of significant words within the previously selected texts. Classical methods of textual feature selection are based on statistical approaches, for example selection based on word frequency (DF) or information gain (IG). These methods works well for large collections of texts and involved pre-processing of the full collection. In our case the frequency of words may vary depending on the selected parts of documents and the size of the resulting collection can be very heterogeneous from one experiment to the other. In order to avoid heterogeneous frequency, we chose a natural language approach where words are tagged and selected according to their syntactic role in the sentence. We use TreeTagger, a tool for annotating text with part-of-speech and lemma information, developed at the Institute for Computational Linguistics of the University of Stuttgart [Schmid 1994, Schmid 1995]. We retain different types of words, depending on the structured feature selection. For experiments K-F and K-all (keywords) we keep nouns, verbs, adjectives, (excluding conjunctions, unknown words, etc.), while for experiments T-PF and T-P (full text), we keep only the nouns to limit the number of features. There is one difficulty with conference names due to their very free and heterogeneous labelling : some teams would use the full name of the conference, others the acronym in various formats (e.g. POPL 03, POPL03, POPL 2003). We therefore built manually a normalized list of all the conference names and matched automatically the form used in the RA with the normalized form. Since conference acronyms are significant and unknown to the tagger, we decided not to use the tagger for this experiment, keeping all the words but stop words (such as proceedings, conference, etc.). Finally, for all experiments, words that are not used at least by two teams are removed. Table 1 summarises the size of data (words) used in each experiment. 3.2 Clustering Method and External Evaluation The third step is clustering of documents in a set of disjoint classes using the vocabulary defined for the five experiments as described above. Our clustering algorithm is based on the partitioning method proposed by [Celeux et al 1989], where the distances between clusters is based on the fre-

5 quency of the words of the selected vocabulary. This approach is equivalent to the k-means algorithm. As for the k-means we represent the clusters by prototypes which summarize the whole information of the document s belonging to each of them. More precisely, if the vocabulary counts p words, each document s is represented by the vector x s = (x 1 s,..., xj s,..., xp s ) where xj s is the number of occurences of word x j in the document s, then the prototype g for a class U i is represented by g i = (gi 1,..., gj i,..., gp i ) with gj i = s U i x j s. Finally, the prototype of each class been fixed, every element is assigned to a class according to its proximity to the prototype. The proximity is measured by a classical distance between distributions (e.g. chi-squared). We evaluate the quality of our automatic clustering by comparing the results with the two sets of themes used by Inria. We call this evaluation external validity, since the clustering process does not involve those themes. For all quantitative evaluations we use two complementary measures: the F-measure and the corrected Rand index. The F-measure proposed by [Jardine and Rijsbergen 1963] combines the precision and recall measures from information retrieval and treats each cluster as if it was the result of a query and each class as if it was the desired answer to that query. For a priori group U i ; and cluster C j, recall(i,j) is equal to n ij /n i. and precision(i,j) is equal to n ij /n.j, where n ij is the number of documents in a group U i and the cluster C j ; n i. the number of documents in a priori group U i ; n.j the number of documents in cluster C j. Then, the F-measure between U i and C j is given by F(i,j)=(2.*recall(i,j)*precision(i,j)/(recall(i,j)+precision(i,j)). The F-measure between a priori partition U and the partition C in K clusters is given by: F = k j=1 n.j n max (F(i, j)) (1) j, where n is the total number of documents in the data set. The corrected Rand (CR) index is defined in [Hubert and Arabie 1985] for comparing two partitions. We remind its definition. Let U = {U 1,...,U i,..., U r } and P = {C 1,...,C j,..., C k } be two partitions of the same data set having respectively r and k clusters. The corrected Rand index is: CR = r i=1 1 2 [ r i=1 k j=1 ( nij 2 ( ni. ) k 2 + j=1 ) ( n ) 1 r 2 i=1 ( n.j ) ( 2 ] n ) 1 r 2 i=1 ( ni. ) k ( n.j ) 2 j=1 2 ( ni. ) k 2 j=1 where ( ) n 2 = n(n 1) 2 and n ij, n i., n.j and n are defined as above. ( n.j ) (2) 2

6 Exp. Nb. of F Rand F Rand F Rand clusters Themes 2003 Themes 2003 subthemes subthemes Themes 2004 Themes 2004 K-F-a K-F-b K-F-c K-all-a K-all-b K-all-c T-P-a T-P-b T-P-c T-PF-a T-PF-b T-PF-c T-C-a T-C-b T-C-c Table 2: Results by external validity To conclude, the F-measure is easier to interpret and can support local analysis (through the F ij ), while the Rand gives a measure of the significance of the results for a given number of clusters. 3.3 Results Analysis Table 2 summarizes our results for different feature selections and different number of clusters (4, 5 and 9). We first analyse results for Themes 2003 and Themes 2004 separately, then compare between the two. For Themes 2003, we get the best results consistently for the two measures and all features when clustering into 4 clusters. One exception is clustering in 9 sub-themes using the text of both presentation and foundation (T-PF-c). The overall best result is obtained with four clusters using presentation and foundation (T-PF-a). A finer analysis using sub-themes (not presented here by lack of space), highlights good mapping between clusters and sub-themes. For Themes 2004, we get good results for clustering in 5 or 4 clusters, with the best results with 5 clusters when using all the keywords. In both cases, the sections about Foundation seem representative of the research domains, either through the full text of those sections or through their attached keywords, for the teams who provided such keywords. Overall, our automatic clustering compares better with Themes 2003 than with Themes 2004, with the exception of using all keywords for creating 5 clusters. There is not much difference when comparing with Themes2003 or Themes 2004 when using the conference names. Somehow disappointing results with conference names may be explained by not using the tagger, leaving us with too many different words (see table 1). We also note that the two measures, F-measure and corrected rand, are consistent trough the experiments: high F-measure scores correspond to good rand values.

7 4 Related Work Currently research in classification and clustering methods for XML or semistructured documents is very active. New document models have been proposed by ([Yi and Dundaresan 2000], [Denoyer et al 2003]) to extend the classical vector model and take into account both the structure and the textual part. It amounts to distinguish words appearing in different types of XML elements in a generic way, while our approach uses the structure to select (manually) the type of elements relevant to a specific mining objective. XML document clustering has been used mostly for visualizing large collections of documents, for example [Guillaume and Murtagh 2000] cluster AML (Astronomical Markup Language) documents based only on their links. [Jianwu and Xiaoou 2002] propose a model similar to [Yi and Dundaresan 2000] but adding in- and out-links to the model, and they use it for clustering rather than classification. [Yoon and Raghavan 2001] also propose a BitCube model for clustering that represents documents based on their epaths (paths of text elements) and textual content. Their focus is on evaluating time performance rather than clustering effectiveness. Another direction is clustering Web documents returned as answers to a query, an alternative to rank lists. [Zamir and Etziono 1998] propose an original algorithm using a suffix tree structure, that is linear in the size of the collection and incremental, an important feature to support online clustering. [Larsen and Aone 1999] compare different text feature extractions, and variants of a linear-time clustering algorithm using random seed selection with center adjustment. 5 Conclusion and Future Work In this paper we have presented a methodology for clustering XML documents and evaluate the results, for different feature selections, in comparison with two existing typologies. Although the analysis is closely related to our specific collection, we believe that the approach can be used in other contexts, for other XML collections where some knowledge of the semantic of the DTD is available. The results show that the quality of clustering strongly depends on the selected document features. In our application, clustering using foundation sections always outperforms clustering using keywords. This conclusion can be turned the other side down, as an indication for the organization that some parts of the Activity Report do not appropriately describe the research domains and that the choice of keywords and research presentation could be improved to carry a stronger message. On more technical aspects, our approach provides a flexible clustering framework where structured features and textual features can be selected indepen-

8 dently, although comparisons should be done with textual feature selection based on statistical approach (tf*idf). Finally we plan to carry further experiences with conference names using an ontology of conferences. A first step would be to build a good classifier able to match incomplete and incorrect conference or journal titles with their normalized forms. Acknowledgements: The authors wish to thank Mihai Jurca, engineer in the AxIS team for his useful support in the feature pre-processing of the data. References [Celeux et al 1989] Celeux G., Diday E., Govaert G., Lechevallier Y., Ralambondrainy, H. (1989): Classification Automatique des Données, Environnement statistique et informatique. Bordas, Paris. [Denoyer et al 2003] Denoyer L., Vittaut J.-N., Allinari P., Brunessaux S., Brunessaux S.: Structured Multimedia Document Classification, in DocEng 03, November 20-22, 2003, Grenoble, France, article/783.pdf. [Despeyroux 2004] Despeyroux Th.: Practical Semantic Analysis of Web Sites and Documents, in proceedings of the 13th World Wide Web Conference (WWW2004) New York City, May (2004). [Guillaume and Murtagh 2000] Guillaume D., Murtagh F.: Clustering of XML documents, in Computer Physics communications 127 (2000) [Hubert and Arabie 1985] Hubert L., Arabie P. (1985), Comparing Partitions, Journal of Classification, Vol. 2, pp [Jardine and Rijsbergen 1963] Jardine, N., van Rijsbergen, C.J.: The use of hierarchic clustering in information retrieval, Information Storage retrieval, 7, pp , [Jianwu and Xiaoou 2002] Jianwu Y., Xiaoou C.: A semi-structured document model for text mining, Journal of Computer Science and Technology archive, Volume 17(5), , May (2002), [Larsen and Aone 1999] Larsen B., Aone C.: Fast and effective text mining using linear-time document clustering, In: Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining, (1999), pp [Schmid 1994] Schmid H.: Probabilistic Part-of-Speech Tagging Using Decision Trees, revised version, original work in Proc. of the International Conference on New Methods in Language Processing, Manchester, UK, pp , (1994), [Schmid 1995] Schmid H.: Improvements in Part-of-Speech Tagging With an Application To German, revised version, original work in Proc. the EACL SIGDAT workshop, Dublin, (1995), [Yang and Pederson 1997] Yang Y., Pederson J.O.: A comparative study on Feature Selection in text categorisation, in Proceedings of the Fourteenth International Conference on Machine Learning, pages , Morgan Kaufmann, (1997). [Yi and Dundaresan 2000] Yi J. and Dundaresan N.: A classifier for semi-structured documents, in Proc. of the 6th International Conference on Knowledge Discovery and Data mining (KDD), (200), [Yoon and Raghavan 2001] Yoon J., Raghavan V.: BitCube: Clustering and Statistical Analysis for XML Documents, Journal of Intelligent Information Systems, (2001). [Zamir and Etziono 1998] Zamir O., Etzioni O.: Web Document clustering: A feasibility demonstration, in ACM Conf SIGIR98, Melbourne, Australia, (1998).